Authority is a decision,
not a model attribute.

An AI agent does not become trustworthy because a dashboard reports a strong average result. It becomes eligible for a wider operating boundary only when an accountable person decides that the evidence is sufficient for that specific change.

That is a small distinction in language and a large distinction in operation.

An agent may successfully handle a bounded task: review low-risk requests, negotiate a defined class of supplier terms, or route standard exceptions. The next question is not simply whether its performance is “good.” The question is whether it may now handle a larger spend range, a new exception class, a different supplier segment, or an action without the prior human checkpoint.

Those are authority decisions. They have an owner, a consequence, an evidence basis and a point in time.

The missing record

Most organisations already retain parts of the story: model or workflow outputs, audit logs and approvals, exception queues, realised outcomes, and later investigations when something goes wrong.

But the decision to change an agent boundary is often reconstructed afterwards from scattered systems and personal recollection. That makes the review slow, makes responsibility ambiguous, and invites a dangerous shortcut: using aggregate performance as if it proved every next expansion is safe.

It does not.

Evidence must be relevant to the boundary being changed. A strong result in routine invoice matching does not, by itself, justify autonomous supplier award. A high success rate in one market does not erase a new compliance constraint in another. Equally, one noisy failure should not automatically freeze an otherwise well-bounded and well-evidenced expansion.

A reviewable boundary decision

A useful decision record should make five things clear:

1. What was the agent allowed to do? The prior operating boundary.
2. What change was considered? A narrower or wider authority, stated concretely.
3. What was known then? Evidence available at the time of decision — not conclusions imported from later outcomes.
4. Who owned the decision? The person or body able to accept the resulting exposure.
5. What happened afterwards? Held separately as outcome evidence for the next decision.

This is not a score that certifies an agent. It is not a way to remove human accountability. It is a method for making accountability operational: the decision owner can see what was relied upon, what was unknown, and why the boundary did or did not move.

The question worth testing

Before an agent’s authority changes, can the decision owner obtain a compact, reviewable reconstruction of the relevant evidence — quickly enough to be useful and honestly enough to be trusted?

If the answer is yes, governance becomes a mechanism for safe operating capacity rather than a retrospective compliance exercise. If the answer is no, the next priority is not broader autonomy. It is understanding what evidence, ownership or process is missing.

EARNED is researching that question. The work is early. It does not provide production autonomy, safety certification or an automated decision about what an agent may do.